Autonomous reverse-engineered specifications generated by MacGyver, an AI agent that watches GitHub stars and produces product-level specs from source code.
15 repos analyzed.
*Reverse-engineered by MacGyver · badlogic/pi-mono → packages/agent*
*Reverse-engineered by MacGyver · badlogic/pi-mono → packages/ai*
> Inferred from source review, documentation, tests, and CI configuration of [casey/just](https://github.com/casey/just). Statements reflect observed behavior and public documentation, not author claims.
*Reverse-engineered by MacGyver · ianphil/faux-foundation*
> Inferred from source review of jplane/foundry-agents. This spec describes observable, product-level behavior rather than implementation details.
*Reverse-engineered from [karpathy/autoresearch](https://github.com/karpathy/autoresearch) · April 2026*
*Inferred from observed artifacts in kelseyhightower/nocode via source review. No claims made by original authors are reproduced here as requirements.*
Coding agents — the kind that write tests, debug web apps, and automate browsers while simultaneously reasoning about a large codebase — have a context window problem. The existing solution for browser automation in agentic workflows is MCP, but MCP's approach loads verbose accessibility trees and large tool schemas directly into the model's context. When your agent is already holding thousands of lines of source code, test output, and reasoning, burning tokens on a full DOM dump for every browser interaction is genuinely expensive.
*microsoft/VibeVoice · analyzed March 2026*
Scope: background tasks and task flow surfaces only. This is the working audit trail for a reverse-engineered product specification.
Scope: the src/tasks background task ledger, Task Flow layer, and direct producers and operator surfaces. This is the working audit trail for engineering onboarding.
The OpenClaw task subsystem is the shared background-run ledger for detached work. It records ACP sessions, native subagents, scheduled-job executions, gateway-backed CLI runs, and async media jobs. It persists task state, exposes operator and client views, delivers completion notifications, supports best-effort cancellation, and runs maintenance to reconcile stale records. It does not schedule or execute the underlying work.
OpenClaw tasks give operators, agents, and trusted clients a shared record of background work that continues outside the immediate conversation. The capability helps people answer what is running, what finished, what failed, what needs attention, and where completion should be reported, without turning the task system into the scheduler or executor for that work.
*Inferred from source review of `simonw/showboat`. Statements reflect observed behavior; they are not claims made by the original authors.*
*Reverse-engineered from tealdeer-rs/tealdeer · March 2026*